Summary

• Aim: Classify AGN light curves as containing a flare or not.
• Use GPs to statistically summarise an AGN light curve.
• Applied GPR to progressively more realistic AGN light curves:
  • idealised light curves injected with synthetic flares, to
  • real ZTF light curves.
• Detected 27 AGN flare candidates out of 9035 Type 1 AGN light curves.
• Effects of observational cadence severely affects detection.
• FPR are reasonable but TPR is very low for “real” flares.

Data

1. Simulated damped random walk
   • n = 10000
   • Regular 10d cadence
   • Uniform uncertainties of 0.1 mag
2. ZTF Public Data Release 6 (August 2021)
   • n = 9035
   • Optical r-band
   • Type 1 AGN from SDSS DR7, ALPAKA catalogue

Method

1. Pre-process the light curve based on the experimental scenario.
2. Apply Gaussian process regression to model the light curve.
3. Plot the fitted the hyperparameters.
4. Compare the location of flaring and non-flaring AGN light curves in the hyperparameter-space.

Scenarios

1. Simulated, uniform cadence, with injected flares.
2. Simulated, subsampled to ZTF cadence, with injected flares.
3. Scenario 2 with added outliers.
4. ZTF light curves with injected flares.
5. ZTF light curves.

Flares, Cadence, and Outliers

• Injected flares
  • Gaussian and Gamma functions.
  • 1 to 2.5 mag, lasting 100 to 1000 days.
  • injected at random after 300d, but before last 300d.
• Subsampling
  • “interpolated” light curves to ZTF cadences.
• Outliers
  • one contiguous pair.
  • 5 std. deviation above that light curve’s variability.

GP fitting

• Used celerite (Foreman-Mackey et al. 2017).
• Matern 3/2 kernel
  • amplitude, \(\sigma\).
  • length scale, \(\rho\).
• No mention of mean function, presumably zero function?
• Use maximum a posteriori point?

Results 1: Injected Flares

Results 2: Subsampled cadence

Results 3: Added outliers

Results 4: ZTF with injected flares

Probability of flare, \(P_\textrm{Flare}\)

• \(\theta\) is indicator variable of whether light curve contains a flare.

\[P(\sigma, \rho, \theta \mid y) \propto P(y \mid \sigma, \rho) \times P(\sigma, \rho \mid \theta) \times P(\theta)\]

• Used MCMC rejection sampling to get samples of \(\theta\).
  • 10000 samples plus 2000 burn-in.
• Arrived at a threshold of \(P_\textrm{Flare} = P(\theta = 1) = 0.1\).

Classification Results 1

Classification Results 2

ZTF AGN Flare candidates

Comments

• Very nice narrative told in text and figures: synthetic to real data.
• Good to see application of GPs as not just interpolating poorly sampled data.
• Would be nice to see more light curves!
• Good discussion of limitations.
• Important reminder that sometimes low FPR is more important than high TPR, e.g., searching through large volumes of data.

Queries

• What were the parameters of the injected Gaussian and Gamma flares?
• Why is the hyperparameter distribution of ZTF light curves bimodal?
• Cutoff \(P_{Flare} > 0.1\) held fixed for all scenarios, even real light curves which seems unreasonable.
• Is there code available?

Statistical Nitpicking

• No justification for structure of GP, i.e., mean, kernel.
• “Recall” not “Retrieval” (aka. true positive rate).
• Not clear what the scatter plots are plotting. Maximum a posteriori estimate?
• All point estimates, no uncertainty intervals anywhere!
• Don’t just connect-the-dots for a gappy time-series. Perhaps use smoothing splines for visualisations.

Performance on real > synthetic

• Synthetic scenarios are not reflective of reality?
• More thought needed on shape of flares and how they are inserted into the original light curve.
• Subsampling might induce features, not just remove them.

Extensions

• Multi-band analysis
• Accommodate truncated flares.
• Model the migration of light curves in the hyperparameter-space.

Preach it brother!